fix: vision_load crash recovery and robust image message handling

agent.py

@@ -503,6 +503,18 @@ async def stream_callback(chunk: str, full: str):
                         PrintStyle(font_color="red", padding=True).print(msg["message"])
                         self.context.log.log(type="warning", content=msg["message"])
                     except Exception as e:
+                        # Check if this is a provider rejection caused by raw image content
+                        if "BadRequestError" in type(e).__name__ or "400" in str(e):
+                            stripped = self._strip_raw_images_from_history()
+                            if stripped:
+                                self.context.log.log(
+                                    type="warning",
+                                    content=f"Stripped {stripped} raw image(s) from history after provider rejection (inner loop)",
+                                )
+                                PrintStyle(font_color="orange", padding=True).print(
+                                    f"Stripped {stripped} raw image(s) from history after provider rejection, retrying..."
+                                )
+                                continue
                         # Retry critical exceptions before failing
                         error_retries = await self.retry_critical_exception(
                             e, error_retries
@@ -520,6 +532,18 @@ async def stream_callback(chunk: str, full: str):
                 error_retries = 0  # reset retry counter on user intervention
                 pass  # just start over
             except Exception as e:
+                # Check if this is a provider rejection caused by raw image content
+                if "BadRequestError" in type(e).__name__ or "400" in str(e):
+                    stripped = self._strip_raw_images_from_history()
+                    if stripped:
+                        self.context.log.log(
+                            type="warning",
+                            content=f"Stripped {stripped} raw image(s) from history after provider rejection (outer loop)",
+                        )
+                        PrintStyle(font_color="orange", padding=True).print(
+                            f"Stripped {stripped} raw image(s) from history after provider rejection, retrying..."
+                        )
+                        continue
                 # Retry critical exceptions before failing
                 error_retries = await self.retry_critical_exception(
                     e, error_retries
@@ -581,6 +605,40 @@ async def prepare_prompt(self, loop_data: LoopData) -> list[BaseMessage]:
 
         return full_prompt
 
+    def _strip_raw_images_from_history(self) -> int:
+        """Strip raw image messages from history after provider rejection.
+
+        When a provider returns 400 BadRequestError for image content,
+        this replaces raw image messages with text previews to prevent
+        permanent crash loops where every subsequent message triggers
+        the same 400 error.
+
+        Returns the count of stripped images.
+        """
+        stripped = 0
+        # Check current topic messages
+        for msg in self.history.current.messages:
+            if history._is_raw_message(msg.content):
+                preview = (
+                    msg.content.get("preview", "<image content removed after provider error>")
+                    if isinstance(msg.content, dict)
+                    else "<image content removed after provider error>"
+                )
+                msg.set_summary(preview)
+                stripped += 1
+        # Check past topics
+        for topic in self.history.topics:
+            for msg in topic.messages:
+                if history._is_raw_message(msg.content):
+                    preview = (
+                        msg.content.get("preview", "<image content removed after provider error>")
+                        if isinstance(msg.content, dict)
+                        else "<image content removed after provider error>"
+                    )
+                    msg.set_summary(preview)
+                    stripped += 1
+        return stripped
+
     async def retry_critical_exception(
         self, e: Exception, error_retries: int, delay: int = 3, max_retries: int = 1
     ) -> int:

python/helpers/history.py

@@ -215,8 +215,20 @@ async def compress_attention(self) -> bool:
         return False
 
     async def summarize_messages(self, messages: list[Message]):
-        # FIXME: vision bytes are sent to utility LLM, send summary instead
-        msg_txt = [m.output_text() for m in messages]
+        # Explicitly replace raw message content (e.g., base64 images) with
+        # text previews to avoid sending binary data to the utility model
+        msg_txt = []
+        for m in messages:
+            if _is_raw_message(m.content):
+                preview = (
+                    m.content.get("preview", "<non-text content>")
+                    if isinstance(m.content, dict)
+                    else "<non-text content>"
+                )
+                label = "ai" if m.ai else "user"
+                msg_txt.append(f"{label}: {preview}")
+            else:
+                msg_txt.append(m.output_text())
         summary = await self.history.agent.call_utility_model(
             system=self.history.agent.read_prompt("fw.topic_summary.sys.md"),
             message=self.history.agent.read_prompt(
@@ -536,6 +548,13 @@ def output_text(messages: list[OutputMessage], ai_label="ai", human_label="human
 
 
 def _merge_outputs(a: MessageContent, b: MessageContent) -> MessageContent:
+    # Guard: convert raw messages to text previews before merging
+    # to prevent invalid mixed-format message structures
+    if _is_raw_message(a):
+        a = _stringify_content(a)
+    if _is_raw_message(b):
+        b = _stringify_content(b)
+
     if isinstance(a, str) and isinstance(b, str):
         return a + "\n" + b
 

python/tools/vision_load.py

@@ -71,11 +71,27 @@ async def after_execution(self, response: Response, **kwargs):
                             "text": "Error processing image " + path,
                         }
                     )
-            # append as raw message content for LLMs with vision tokens estimate
-            msg = history.RawMessage(raw_content=content, preview="<Base64 encoded image data>")
-            self.agent.hist_add_message(
-                False, content=msg, tokens=TOKENS_ESTIMATE * len(content)
-            )
+            try:
+                # append as raw message content for LLMs with vision tokens estimate
+                msg = history.RawMessage(
+                    raw_content=content,
+                    preview=f"<Base64 encoded image data ({len(content)} image(s))>",
+                )
+                self.agent.hist_add_message(
+                    False, content=msg, tokens=TOKENS_ESTIMATE * len(content)
+                )
+            except Exception as e:
+                # Graceful degradation: add text description instead of raw image
+                # to prevent corrupted image data from persisting in chat history
+                PrintStyle().error(f"Error adding image to history: {e}")
+                self.agent.context.log.log(
+                    "warning", f"Error adding image to history: {e}"
+                )
+                fallback = (
+                    f"{len(self.images_dict)} image(s) loaded but could not be "
+                    f"added to context: {e}"
+                )
+                self.agent.hist_add_tool_result(self.name, fallback)
         else:
             self.agent.hist_add_tool_result(self.name, "No images processed")